PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-Aided Design for Manufacturing: A Multi-Agent System for Intent-Preserving Redesign of CAD for Improved Manufacturability

Kojo Welbeck, Xiangyu Shi, Zahra Sadeghi, Qi Zhu, Ping Guo

arXiv:2609.05559v1cs.CEcs.CVcs.MA

TL;DR

AI-generated CAD fidelity degrades as parts grow more complex, challenging CAD generation for engineering use. The paper proposes intent-preserving redesign through sequentially verified edits and reports lower mean chamfer distance than same-tool chain-of-thought single-agent baselines on a 46-part benchmark.

  • Problem

    AI-generated CAD fidelity declines as parts grow more complex, limiting the usefulness of generated models for downstream engineering reasoning.

    Source: §1 · p. 2

  • Method

    A multimodal LLM-driven system reviews an existing CAD part and applies proposed manufacturability improvements as individually compiled and visually checked edits to its CadQuery program.

    Source: §1 · p. 2, CAD Modifier · p. 7

  • Results

    On a 46-part benchmark, the CAD Modifier attains lower mean chamfer distance than same-tool chain-of-thought single-agent baselines; removing visual review raises raw mean distance from 3.496 to 11.495.

    Source: Ablation Study · p. 15, §7 · p. 19

  • Takeaways & Limitations

    Chaining verified edits provides a way to build more complex designs while maintaining design intent at each step.

    Source: §1 · p. 2

    Intent preservation currently lacks an operational definition and rests on visual inspection, so substantiation requires a pre-declared metric.

    Source: §6 · p. 19

Abstract

from arXiv · show

We introduce autonomous, intent-preserving Design for Manufacturing (DFM) redesign of CAD parts: given an engineer's CAD model, the method returns a variant that is easier to manufacture without losing its design intent. Generating such a redesign in a single shot is unreliable, since CAD fidelity degrades as parts grow complex; we instead produce it as a sequence of individually verified design transitions. Our DFM-Redesign pipeline realizes this with two coupled agent subsystems driven by a pretrained multimodal LLM: a DFM Reviewer that inspects the current design and proposes one intent-preserving manufacturability improvement at a time, and a CAD Modifier that executes each proposal as an edit to the part's CadQuery program. The CAD Modifier closes a verification loop, compiling every candidate edit and visually checking it against the intended change from multi-view renderings, then re-generating or re-instructing until the edit is accepted or abandoned. Iterating review and verified modification compounds edits into parts more complex than one-shot generators reliably produce, preserves the original intent at each step, and requires no fine-tuning. On a 46-part benchmark scored by chamfer distance to reference geometries, the CAD Modifier reproduces target parts more accurately on average than chain-of-thought single agents given the same tools, and ablations isolate the contributions of the visual review loop and of captioning the design state before each edit. A centrifugal pump casing built from 32 chained transitions illustrates the complexity reachable by compounding verified edits. This is a preliminary report: evaluation of the full review-and-redesign loop, including manufacturability gain and an operational measure of intent preservation, is ongoing.

1 Introduction

AI-generated CAD can accelerate product development, but fidelity declines as parts become more complex. This work proposes intent-preserving DFM redesign through individually verified transitions in an autonomous multi-agent pipeline.

1 Introduction

  • For the badge example, the pipeline engraves raised “A C” lettering 2 units into the surface, then adds 1-unit-radius fillets to concave character corners.The final part is described as more agreeable to CNC machining while retaining the badge design intent.Source: Figure 1 · p. 2
  • CAD fidelity degrades as parts grow more complex, limiting the usefulness of generated models for downstream engineering reasoning.Automated CAD generation could support product development from ideation through manufacturing specifications and broaden access beyond trained engineers.Source: §1 · pp. 1–2
  • The method treats redesign as a sequence of individually verified edits, rather than generating a whole part in one pass.The authors argue that compounding verified transitions lets a general-purpose multimodal LLM build more complex parts while preserving design intent.Source: §1 · p. 2
  • The authors identify transition-wise generation, an autonomous DFM-Redesign pipeline, and evaluation protocols and datasets as the work’s three contributions.The transition method uses natural-language instructions and image-based feedback loops.Source: §1 · p. 3
  • DFM-Redesign uses an autonomous multi-agent system to review a CAD part and improve its CNC manufacturability while retaining its original intent.The system uses a general-purpose multimodal LLM without fine-tuning.Source: §1 · p. 3

2 Relevant Work

The paper builds on CAD’s editable textual representations and the link between text and visual geometry. It distinguishes its controlled, verified transitions from prior piece-wise and direct-generation approaches.

2 Relevant Work

  • The paper positions intent-preserving DFM redesign as a concrete, training-free multi-agent system for manufacturability, contrasting it with earlier broad conceptions and related multi-agent work.It cites Encube (Encube, 2025) and AgentsCAD (George et al., 2026).Source: §2 · p. 3

CAD as a Precise Geometric Language

  • CadQuery provides an open-source Python codebase for parametric 3D modeling, with scripts that can be read, run, and modified by both an LLM and a human designer.The paper also uses CadQuery’s export capabilities for interoperability and image generation.Source: CAD as a Precise Geometric Language · pp. 3–4

Using Textual Visual Equivalence

  • The system places multi-view renderings inside a generation-time verification loop, unlike text-to-CAD annotation pipelines that use images to caption training data.Text and visual geometry provide complementary information about a CAD part’s construction sequence and state.Source: Using Textual Visual Equivalence · p. 4

From Direct Generation to Piece-wise Generation

  • The paper chains goal-directed, individually verified edits from an existing design, unlike prior methods that generate fixed outputs, make uncontrolled edits, or require task-specific training data.The authors argue that chaining transitions supports increasingly complex models; the piece-wise approach also allows review at each step.Source: From Direct Generation to Piece-wise… · pp. 4–5

3 Methodology

The methodology couples a DFM Reviewer, which proposes context-aware manufacturing transitions, with a CAD Modifier that compiles and visually checks each edit. Accepted edits become the next design state, allowing the part to evolve through sequential transitions.

3 Methodology

  • The pipeline takes the current CAD script, rendered views, and textual part description, then returns an updated CadQuery script for the next design state.The CAD Modifier is downstream of the reviewer and is the central workhorse of the end-to-end system.Source: §3 · p. 5, CAD Modifier · p. 5, Figure 2 · p. 6

CadQuery and Multi-view Images

  • CadQuery scripts represent the part, while eight preset rendered perspectives support visual inspection and Boolean-derived transition images highlight geometric differences.The preset views include isometric, front, rear, right, left, top, bottom, and opposite isometric.Source: CadQuery and Multi-view Images · p. 5

CAD Modifier

  • The CAD Modifier executes a requested transition using three agents: script generation and compilation, visual validation, and transition re-instruction after repeated validation failures.The accepted script and updated part description can serve as the next design state; failed transitions may be abandoned after a configured retry limit.Source: CAD Modifier · p. 7

DFM Reviewer

  • The DFM Reviewer proposes one context-aware transition at a time from the current design, and prompts restrict recommendations to intent-preserving CNC machining changes or none when unnecessary.The reviewer examines the current script and multi-view images, returning a recommendation with a title and rationale.Source: DFM Reviewer · pp. 7–8

4 Demonstration

The demonstration follows a badge redesign through accepted edits and a failed transition that returns the system to its last accepted state. A second example shows manufacturing-oriented changes to a slotted circular-boss part.

Demonstration scope

  • The section presents sample redesigns anecdotally before the systematic evaluation and benchmarking results.It also previews additional processed samples and their transition titles.Source: §4 · p. 8

Badge redesign

  • Two badge transitions—engraving raised text and filleting internal letter corners—were accepted, while a compiler-valid chamfer transition was rejected and abandoned after retries.The system reverted to Design State 3, the last accepted state, as the final design.Source: §4 · pp. 8, 10, Figure 3 · p. 9

Second example

  • For the slotted circular-boss part, the redesign replaces horizontal convex fillets with 45° chamfers and adds fillets to internal vertical slot corners.The figure shows the original, redesigned part, and highlighted difference; Table 1 lists the transition instructions and rationales.Source: Figure 4 · p. 10, §4 · p. 10

Badge redesign

  • The failed badge edit exposed two pitfalls: a potentially unnecessary recommendation and visual review that rejected an initial attempt it should have accepted.The recommendation would have added production steps; re-instruction changed the intent to exclude the ring's inner edge.Source: §4 · p. 10

5 Evaluation

Across a 46-part benchmark, the orchestrated CAD Modifier attained lower mean chamfer distance than the four single-agent benchmarks, though single agents were competitive on parts that went well. The evaluation also shows how chained edits build complex parts, how generation varies between runs, and how individual components affect outcomes.

Complex Part Demonstration

  • The CAD Modifier built the most complex demonstrated part, a centrifugal pump casing, through 32 sequentially applied transitions based on a manually supervised LLM annotation pipeline.The reference geometry came from GrabCAD; the transition order was predetermined manually.Source: Figure 5 · p. 12, Complex Part Demonstration · p. 12

Variability in Part Generation

  • Identical transition sequences could yield different parts; in one inaccurate pump-casing run, acceptance of a misoriented base cylinder led to misapplied relative features such as taper ribbing.The review-loop gating decision and interpretation of the current design and transition descriptions are cited as sources of variability.Source: Variability in Part Generation · pp. 12–13, Figure 6 · p. 13

Evaluation Dataset and the Chamfer Distance Metric

  • The evaluation dataset contains 46 parts with image sets, STEP references, and manually reviewed transition descriptions, assessed by normalized chamfer distance after global and local point-cloud registration.Descriptions began each build in the target orientation, while the distance assessment relaxed that requirement through registration.Source: Evaluation Dataset and the Chamfer… · p. 13

Ablation Study

  • Variant B had lower chamfer-distance averages than Variant A both for 12 runs where description revision triggered and 34 where it did not, while Variant E and B were nearly identical on 35 parts.Seven Variant E samples with worse outcomes skewed its averages; the authors say the adverse effect of description revision was not straightforward to substantiate.Source: Ablation Study · pp. 14, 15, 17, Table 3 · p. 14, Table 5 · p. 17
  • Variant B with pre-transition captioning more often produced lower chamfer distances than Variant C without captioning, with examples of more accurate parts.The captions provide contextual, summarized understanding of the script, images, and transition descriptions.Source: Ablation Study · p. 15, Table 4 · p. 15
  • Removing the visual review loop raised raw mean chamfer distance from 3.496 to 11.495 and anomaly-excluded mean from 2.181 to 4.087, with more frequent and larger accuracy spikes.Without the loop, positioning and orientation errors went unchecked and compounded across transitions.Source: Ablation Study · p. 15, Figure 8 · p. 16
  • Variant F using Gemini 3.5 Flash had lower chamfer distances across almost all parts than Variant B using Gemini 3 Flash Preview, with fewer compilation errors, retries, and less time.The authors report that changing the model moved summary statistics less than removing orchestration.Source: Ablation Study · p. 17, Table 6 · p. 17, Benchmarking · p. 18

Benchmarking

  • 3.496 and 3.634 versus 4.474 to 47.439 raw mean chamfer distance, and 2.181 and 1.846 versus 2.696 to 4.926 with anomalies excluded, favor the orchestrated variants over all four single-agent benchmarks.One single-agent configuration had the lowest median, 1.384; two single-agent configurations had standard deviations of 194.168 and 208.186.Source: Table 7 · p. 18, Benchmarking · p. 18

6 Limitations and Ongoing Work

The current report leaves several claims and comparisons unverified: it evaluates the CAD Modifier rather than the complete redesign pipeline, and key measures and validation studies remain outstanding.

Evaluation scope

  • The evaluation characterizes the CAD Modifier, not the complete DFM-Redesign pipeline.The full review-and-redesign evaluation is still ongoing.Source: §6 · p. 18

Statistical strength

  • Each variant was run once across 46 parts, without significance testing; inspected outliers also make some subsets non-comparable.Repeated runs with a uniform, pre-declared outlier rule and paired significance tests are in progress.Source: §6 · p. 18

Manufacturability measurement

  • Manufacturability improvement over the original parts has not been measured.The intended measures include machining cost and time deltas, setups and tool changes, and removal of hard-to-machine features.Source: §6 · pp. 18–19

Intent preservation

  • Design-intent preservation currently rests on visual inspection rather than a pre-declared operational metric.The proposed metric should address geometric similarity, retained functional features, and dimensional deviation within a stated tolerance.Source: §6 · p. 19

Reachable complexity

  • The pump casing demonstrates chained-transition complexity, but transition success has not been plotted against part complexity.The paper identifies that sweep as the direct evidence needed for its central methodological claim.Source: §6 · p. 19

Evaluators and baselines

  • Fitness-for-purpose lacks human or independent-model evaluation, and comparisons with manual DFM review, established tools, and prior edit-based methods remain outstanding.The proposed evaluator should come from a different model family than the system under test.Source: §6 · p. 19

Benchmark configurations

  • Table 7's four benchmark runs are not individually labeled by model and transition-batching strategy, and one row awaits source-data re-verification.The comparison should be read as orchestrated pipeline versus single agent, not as a comparison among individual rows.Source: §6 · p. 19

Dataset release

  • The authors intend to release the 46-part evaluation set and its complexity spread.The planned release includes multi-view images, STEP files, and manually reviewed transition descriptions.Source: §6 · p. 19

7 Conclusion

The conclusion presents a training-free, multi-agent pipeline that redesigns CAD through individually verified edits. Its CAD Modifier outperforms same-tool chain-of-thought single-agent baselines on mean chamfer distance, while evaluation of the full redesign loop remains ongoing.

Conclusion

  • The pipeline compounds individually verified edits, checking each by compilation and multi-view visual inspection to limit the fidelity loss associated with one-shot X-to-CAD generation.The system is training-free and uses a multimodal model for manufacturability review and redesign.Source: §7 · p. 19
  • On a 46-part benchmark, the CAD Modifier achieves a lower mean chamfer distance than chain-of-thought single-agent baselines equipped with the same tooling.The reported ablations isolate contributions from the visual review loop and pre-transition captioning.Source: §7 · p. 19
  • The approach can be retargeted to stronger general-purpose models without fine-tuning, while end-to-end evaluation of the review-and-redesign loop is ongoing.The ongoing evaluation includes manufacturability gain and an operational measure of intent preservation.Source: §7 · p. 19
Loading 2609.05559v1…